From: AAAI Technical Report WS-00-09. Compilation copyright © 2000, AAAI (www.aaai.org). All rights reserved. Control of Multiple Small Surveillance Robots at AAAI2000 Paul E. Rybski, Dean F. Hougen, Sascha A. Stoeter, Maria Gini and Nikolaos Papanikolopoulos Center for Distributed Robotics, Departmentof ComputerScience and Engineering University of Minnesota {rybski,hougen,stoeter,gini, npapas}@cs,umn. edu Abstract Wedescribe Ourexperiencesdemonstratinga teamof miniature robots engagedin a surveillance missionat the AAAI 2000MobileRobotsExhibition.Wepresent the architecture and the mainfeatures of our systemand discuss someof the environmental factors that affect its performance. Introduction Reconnaissanceand surveillance tasks can benefit from the use of multiple small yet highly capable robots. Suchrobots must be easily deployable and able to moveefficiently yet traverse obstacles or uneventerrain. Theymust be able to sense their environment,act on their sensing, and report their findings. Theymust be able to be controlled in a coordinated manner. Wehave developed a set of extremely small robotic systems, called scouts (Hougenet al. 2000a), which were designed specifically to meet these requirements. The scout robot, shownin Figure 1, is a cylindrical robot 11 cmin length and 4 cmin diameter. Scouts have two modesof locomotion designed to transport themover different kinds of terrain and obstacles. In the first mode,the scouts use their wheelsto roll over smooth surfaces (even upa 20 degree slope). Whenconfronted with an obstacle taller that itself, the scout employsits second locomotion mode, the jump. The spring-loaded tail of the scout is compressedand released to propel the robot over objects upwardsof 20 cm in height. Figure 2 showsa scout jumpingover a barrier. Scouts carry a small video cameraand video transmitter whichthey use to capture information about their environment. They can also transmit and receive digital information over a separate RF communicationssystem which uses a packet-based communicationsprotocol. The small size of the scouts provides manyadvantages. Theyare inexpensive and easily transportable, which makes them ideal for use in large teams. This allows themto be present throughouta wide area, forminga mobilesensor network.It also allows individual scouts to be expendablewithout jeopardizingan entire mission, scouts are well suited to clandestine operations since they can be concealed easily. Copyright(~) 2000,American Associationfor Artificial Intelligence(www.aaai.org). All rights reserved. 33 Figure 1: Twoscout robots (shown next to a quarter for scale). Q2000by ACM,appeared in Rybski et al. (2000). Their small size and cylindrical shape also allows themto be deployed by launching or throwing by hand. Related Work Automatic security and surveillance systems using cameras and other sensors are becoming more common.These typically use sensors in fixed locations, either connected ad hoc or, increasingly, through the shared communication lines of "intelligent buildings" (Porteous 1995). These maybe portable to allow for rapid deployment(Pritchard et al. 1998) but still require humanintervention to reposition whennecessary. This shortcomingis exacerbatedin cases in whichthe surveillance team does not have full control of the area to be investigated, as happensin manylaw-enforcement scenarios. Static sensors have another disadvantage. They do not provide adaptability to changes in the environment or in the task. In case of poor data quality, for instance, we mightwant the agent to movecloser to its target in order to senseit better. Mobile robotics can overcomethese problems by giving the sensor wheels and autonomy. This research has traditionally focused on single, large, independentrobots designed to replace a single humansecurity guard as he makes his rounds (Kajiwara et al. 1985). Such systems are now available commerciallyand are in place in, for example,factory, warehouse, and hospital settings (Kochan1997), and Figure 2: A scout jumpingover a barrier (sequence starts from the upper left corner). ©2000by IEEE,appeared in Hougenet aL (2000a). research continues along these lines (Dillmannet al. 1995; Osipov, Kemurdjian, & Safonov 1996). However, the single mobile agent is unable to be manyin places at once-one of the reasons whysecurity systems were initially developed. Further, large mobilerobots are unable to conceal themselves, which they mayneed to do in hostile or covert operations. Theymayalso be too large to exploretight areas. Multiple robots often can do tasks that a single robot would not be able to do, or can do them faster and more reliably, as described in the extensive survey by Cao, Fukunaga, & Kahng(1997). Tasks traditionally accomplished with multiple robots are foraging (Matari6 1997), formation marching (Balch & Arkin 1998), map making (Burgard et aL 2000), janitorial service (Parker 1996), security (Everett & Gage1999), and exploration (Hougenet aL 2000b). Traditionally, mobile robots have ranged in size from roughly dog-sized to somewhatlarger than a human. For small robots, Lego blocks and microprocessor boards, such as the Handyboard(Martin 1998), are often used to quickly prototype robots. Wehave designed multiple Lego-based robots and used themfor a variety of navigation tasks (Rybski et al. 1998). Recently a significant interest has arisen in designing even smaller mobile robots for exploration and reconnaissance. These include the popular Kheperarobot (Mondada, Franzi, & Ienne 1993) and other prototypes, such as ALICE(Caprari et al. 1998) and our ownscouts (Hougenet aL 2000a), developed by various research groups and not yet commerciallyavailable. The major challenge in designing miniature robots is in fitting all the mechanicalparts and electronics into a limited volume and in designing an adequate method of locomotion. Most miniature robots have wheels (Baumgartner et al. 1998), but some can jump (Halme, Sch6nberg, & Wang 1996), roll (Chemel,Mutschler, &Schempf1999), fly (Wu, Schultz, & Agah 1999), or swim (Fukuda, Kawamoto, Shimojima1996). Our miniature robots have two rolling wheelsbut they can also jumpto go over small obstacles, as shownin Figure 2. 34 Dueto their small size, most miniature robots use proxy processing, as in Inaba et al. (1996), and communicate via radio link with the unit wherethe computationis done. Limited communication bandwidth becomes a problem when robots have to transmit large amountsof data, such as live video, and whenmanyrobots need to share the bandwidth. Power consumption is another major problem which limits the mobility, ability to communicate,and long term survivability of miniaturerobots. Scout Control Architecture The scout’s small size and limited powersupply greatly restricts the kinds of on-boardprocessingthat can take place. The only computations done on the scout’s two 8-bit microprocessors are what’s necessary for communicationsand ¯ actuation. All high-level decision makingand video image processing must be relegated to an off-board workstation or a humanteleoperator. Autonomous decision prgcess, such as planners, reactive behaviors, or teleoperation controls, send low-level actuator commandsto the scouts through an RF transceiver. Each scout is assigned a unique network ID which allows packets to be routed to the correct robot. By interleaving the transmission of packets destined for different robots, multiple scouts can be controlled simultaneously. The current bottleneck in the systemis the numberof packets per second that can be transmitted. The frequency at which commands can be broadcast is currently between5 to 8 per second. The limiting factor is the speedat whichthe scout can decodethe messagesfrom the radio. Scout video data is captured by a video receiver. This can be either displayed on a monitoror fed into a digitizer. Because the video is a continuous analog stream, only one robot can broadcast at any instant. Signals broadcast simultaneously from multiple robots would interfere with each other, makingboth signals useless. Scouts runningin parallel mustinterleave their video transmissions and limit their transmissionsto short bursts. Wehave developed a distributed, CORBA-based (Group The scout knowsthat it should stop whenits camerais either pressed up against a dark object, or the scout is in shadows.Scout motionin this behavior is continuousand the scout does not check to see that it has movedby computing differences betweenframes. Other kinds of motion makesuse of this check because the scout does not always receive the commands sent for it due to RFinterference. Framedifferencing is not helpful in this behaviorbecause the scout is unable to movevery quickly. The difference between subsequent frames captured during forward motion is minimal,makingit very difficult for the robot to detect forward motion. 1998) architecture to coordinate all of the individual hardware resources in our system. Access to robotic hardware and other computationalresources is controlled through processes called RESOURCE CONTROLLERS(or RCS). Every physical resource is given its ownRCprocess to manageit. Behaviors and decision processes connect to these RCSto send commandsand receive data from them. Behavior processes running on different computerscan access all of the RCswithout even being aware that they maybe running on different computers. Scout Behaviors The scouts are capable of operating autonomouslythrough the use of simple behaviors. The only environmentalsensor available to the scout is its video camera, the use of which presents several problems. First, the scout’s proximity to the floor severely restricts the area it can view at a time. Secondly, since the video is broadcast over an RF link to a larger robot or a workstation for processing, the quality of the received video often degrades due to of range limitations, proximityof objects that interfere with transmission, and poor orientation of the antennas. Figure 3 shows an example image received from the scout’s camera. The RF noise degradingthis imageincreases the difficulty of distinguishing the objects from the background. Detect Motion:Detecting movingobjects is accomplished using frame differencing. Oncethe scout has been placed in a single location, it mustdeterminethe quality of the video and set a threshold to filter out RFnoise. This is accomplished by doing image differencing on a stream of video and increasing the difference threshold until RF noise is filtered out. Thescout then subtracts sequential images in the video stream and determines whether the scene changes at all (caused by movement in the image). Demonstrationsat the Exhibition At the AAAI2000 Mobile Robot Exhibition, we demonstrated several aspects of our distributed scout control system on two networked laptop computers. The first laptop ran Linux and handled all of the scout control programs. The second ran Windowsand handled the video processing hardware and software. The scout’s command radio was connected to the Linux laptop while a video receiver was connected to the Windowslaptop. In addition to the autonomousbehaviors defined above, a simple teleoperation client was used which allowed a humanoperator to drive a scout and operate its sensor payload and jumping mechanism. Figure 4 illustrates the implementationof the scout control system. Linux Laptop Windows Laptop [ [Autonomous ConsoleTele°perati°n [Behavior Ethemet Figure 3: The worldfrom the scout’s point of view. Here the scout is viewinga lab benchand twochairs at a range of 2m. ©2000by ACM,appeared in Rybski et al. (2000) PCMCIA Framegra~ A numberof different visually-based behaviors have been implemented;see Rybskiet al. (2000) for the full list. The behaviors demonstrated at the AAAI2000 Mobile Robot Exhibition were: SVideo ~ Drive TowardsGoal: Identifying a dark area to movetowardsis a simple matter of scanning across the image at a fixed level on or about the level of the horizon and determining the horizontal position of the darkest area in the image. The mean pixel values in a set of overlapping windowsin the image are determined. The scout selects the darkest windowand drives in that direction. ~ ~ "~ Signal Transmitted Commands H _ . Scout Robots Figure 4: Diagramof the hardwareand software architecture used to control the scout robots. 35 Several RCS were used by the behaviors and teleoperation consoles to control the scouts. The first, and arguably the mostimportant, RCis in charge of controlling access to the radio. The user interface consoles and behaviors have to connect to the radio RC’s CORBA interface to send packet requests to it. The radio RCsupports two kinds of commandpacket requests. In the first request type, the decision process must explicitly instruct the radio to send each command.If commands are sent to the radio faster than they can be transmitted, they are stored in a queueuntil the radio can handle them. In the second request type, the decision process gives the radio a single command and instructs it to repeat it as often as it can. This allows the systemto keep inter-process communicationsto a minimum.If multiple decision processes set up repeated commands,the radio will schedule transmission of each in a round-robin fashion. If the queue of non-repeated commandshas pending requests, the queuewill also be addedto the schedule. The second RCis the framegrabberprocess, which is required for autonomousbehaviors. Images are captured by the framegrabber on the Windowslaptop and are processed based on the requests of the decision process. Several different image operators were implemented,including frame differencing, connectedregion extraction and various statistical operators. To cut downon the amountof networkoverhead, all image processing was done by the framegrabberRC and only the processed image information was passed back to the decision process. At the exhibition, the scout robots were demonstratedin both teleoperated and autonomousmodes. In the teleoperation demo, we showed the rolling and jumping mobility modesof the robot, as well as demonstratingits usefulness as a remote camera. In the autonomousdemos, the scouts served as motion detectors by transmitting video by doing frame differencing operations and notifying a user’s console whenevermotion was perceived in the image frame. Wealso demonstratedthe scout’s ability to servo to a dark object. For this behavior, imageswere analyzed to find the darkest meanvalue in the robot’s visual field. The robot was commandedto drive itself towards the darkest objects and keep themcenteredin its visual field. Discussion & Lessons Learned Because the scout robots are completely dependent on the two RF communicationschannels for operation, any interference on those frequencies can seriously degrade performance. Because the commandRF channel is packet-based (as opposedto streaming), it is somewhatresilient to local interference. The video channel, on the other hand, transmits continuousanalog data whichis very sensitive to interference. Corruption in the video signal can makeit nearly unusable by autonomouscontrol processes. The frequency that we use for our video transmitters is the 900MHz industrial/medical band and is, unfortunately, extremelypopular. Several other groupsat the exhibition used this frequencyto transmit video (most notably the teleoperated blimp used by the BotBall competition). Wehad to work out a scheduling arrangementwith the organizers of the exhibition to share the communicationsband. 36 The small physical size and battery supply of our robots severely limits the kind of RFtransceivers that we can use. Weare also severely limited in the kinds of video capture devices that wecan use. In our current design of the robots, digital cameraswere not an option becauseof lack of available hardwaresmall enoughto fit into our robot as well as lack of computationalpowerto process an image even if we had such a camera. Moreattention will have to be paid to this issue in future scout designs. Onesolution will be to sacrifice robot size in order to include additional or more powerful computational resources. Somesort of a spreadspectrum or frequency-hopping radio system would be extremely useful as well. Not only wouldthe higher frequencies involvedallow for faster data transmission(critical for the transmission of video data), the ability to changefrequencies would allow more robots to operate in the same area and wouldbe morerobust to interference. Wehope to be able to address someof these issues in the future as advances in technology make better and smaller equipmentmoreavailable. Until this happens, we are focusing our efforts on algorithmsfor better video signal processing as well as decision processes which take the inherent fragility of the RF communicationschannels into account whencontrolling the robots. Acknowledgements Wewould like to acknowledge the University of Minnesota’s Doctoral Dissertation Fellowship programfor partial support of this research and the AmericanAssociation for Artificial Intelligence for the AAAIRobot Competition and Exhibition Scholarship that allowedus to bring our robots to AAAI2000 in Austin, TX. This material is based upon work supported by the Defense AdvancedResearch Projects Agency, Microsystems TechnologyOffice (Distributed Robotics), ARPA Order No. G155, Program Code No. 8H20, Issued by DARPA/CMD under Contract #MDA972-98-C-0008. References Balch, T. R., and Arkin, R. C. 1998. Behavior-basedformation control for multiagent robot teams. IEEETransactions on Robotics and Automation 14(6):926-939. Baumgartner, E.; Wilcox, B.; Welch, R.; and Jones, R. 1998. Mobility performanceof a small-bodyrover. In Int’l Symposiumon Robotics with Applications, WorldAutomation Congress. Burgard, W.; Moors, M.; Fox, D.; Simmons, R.; and Thrun., S. 2000. Collaborative multi-robot exploration. In Proc. of the IEEElnt’l Conferenceon Robotics and Automation, 476-481. Cao, Y.; Fukunaga, A.; and Kahng, A. 1997. Cooperative mobile robotics: Antecedents and directions. Autonomous Robots 4(1):7-27. Caprari, G.; Balmer,E; Piguet, R.; and Siegwart, R. 1998. The autonomousmicro robot ALICE:A platform for scientific and commercial applications. In Proc. of 1998 lnt’l Symposium on Micromechatronics and HumanScience (MHS’98). Chemel, B.; Mutschler, E.; and Schempf, H. 1999. Cyclops: Miniature robotic reconnaissance system. In Proc. of the IEEEInt’l Conferenceon Robotics and Automation, 2298-2303. Dillmann, R.; Kaiser, M.; Wallner, E; and Weckesser, P. 1995. PRIAMOS:An advanced mobile system for service, inspection, and surveillance tasks. In Modellingand Planning for Sensor BasedIntelligent RobotSystems, volume21 of Series in MachinePerception and Artificial Intelligence. Singapore:WorldScientific. Everett, H. R., and Gage, D.W. 1999. From laboratory to warehouse: Security robots meet the real world. Int’l Journal of Robotics Research 18(7):760-768. Fukuda, T.; Kawamoto,A.; and Shimojima, K. 1996. Acquisition of swimmingmotion by RBFfuzzy neuro with unsupervised learning. In ALIFEV, 31-37. Group, O.M. 1998. The CommonObject Request Broker: Architecture and Specification. Object Management Group. Halme, A.; Sch6nberg, T.; and Wang, Y. 1996. Motion control of a spherical mobilerobot. In 4th lnt’l Workshop on Advanced Motion Control. Hougen, D. E; Benjaafar, S.; Bonney, J. C.; Budenske, J. R.; Dvorak, M.; Gini, M.; Krantz, D. G.; Li, P. Y.; Malver, E; Nelson, B.; Papanikolopoulos, N.; Rybski, P. E.; Stoeter, S. A.; Voyles, R.; and Yesin, K. B. 2000a. A miniature robotic systemfor reconnaissance and surveillance. In Proc. of the IEEElnt’l Conferenceon Robotics and Automation, 501-507. Hougen, D. E; Erickson, M. D.; Rybski, P. E.; Stoeter, S. A.; Gini, M.; and Papanikolopoulos, N. 2000b. Autonomousmobile robots and distributed exploratory missions. In lnt’l Symposium on Distributed Autonomous Robotic Systems. Inaba, M.; Kagami, S.; Kanechiro, E; Takeda, K.; Tetsushi, O.; and Inoue, H. 1996. Vision-based adaptive and interactive behaviors in mechanical animals using the remote-brained approach. Robotics and AutonomousSystems 17:35-52. Kajiwara, T.; Yamaguchi,J.; Kanayama,Y.; Yuta, S.; and Iijima, J. 1985. Developmentof a mobile robot for security guard. In Proceedingsof the 15th Int’l Symposiumon Industrial Robots, volume1, 271-278. Kochan, A. 1997. HelpMateto ease hospital delivery and collection.tasks, and assist with security. Industrial Robot 24(3):226--228. Martin, E G. 1998. The HandyBoardTechnical Reference. MITMedia Lab, Cambridge, MA. Matari6, M. 1997. Behavior-based control: Examplesfrom navigation, learning, and groupbehavior. Journal of Experimental andTheoretical Artificial Intelligence 9(2-3):323336. Mondada,E; Franzi, E.; and Ienne, P. 1993. Mobilerobot miniaturisation: A tool for investigation in control algorithms. In ExperimentalRoboticsI11, Proc. of the 3rd Int’l 37 Symposium on Experimental Robotics, 501-513. Kyoto, Japan: Springer Verlag, London. Osipov, A.; Kemurdjian, V.; and Safonov, B. 1996. Mobile robots for security. In Proceedingsof the 1996 2nd Specialty Conference on Robotics for Challenging Environments, RCE-H,290-295. Parker, L. E. 1996. Onthe design of behavior-basedmultirobot teams. Journal of AdvancedRobotics 10(6):547-578. Porteous, J. 1995. Intelligent buildings and their effect on the security industry. In Sanson, L. D., ed., IEEElnt’l CarnahanConference on Security Technology, 186--188. Pritchard, D.; White, R.; Adams,D.; Krause, E.; Fox, E.; Ladd, M.; Heintzleman,R.; Sprauer, P.; and MacEachin,J. 1998. Test and evaluation of panoramicimaging security sensor for force protection and facility security. In IEEE lnt’l CarnahanConference on Security Technology, 190195. Larry D. Sanson, ed. Rybski, P. E.; Larson, A.; Lindahl, M.; and Gini, M. 1998. Performanceevaluation of multiple robots in a search and retrieval task. In Workshopon Artificial Intelligence and Manufacturing, 153-160. Rybski, P. E.; Stoeter, S. A.; Erickson, M. D.; Gini, M.; Hougen,D. E; and Papanikolopoulos, N. 2000. A team of robotic agents for surveillance. In Proc. of the lnt’l Conf. on AutonomousAgents, 9-16. Wu, A. S.; Schultz, A. C.; and Agah, A. 1999. Evolving control for distributed micro air vehicles. In Proc. IEEE Int’l Symposiumon ComputationalIntelligence in Robotics and Automation.